PaperlayerView exact revision on arXiv

Source-linked AI summary

CA+: Cognition Augmented Counselor Agent Framework for Long-term Dynamic Client Engagement

Yuanrong Tang, Yu Kang, Yifan Wang, Tianhong Wang, Chen Zhong, Jiangtao Gong

arXiv:2503.21365v1cs.HC

TL;DR

AI counseling systems face difficulty sustaining long-term client engagement across evolving therapeutic contexts. The paper proposes CA+, a cognition-augmented counselor agent framework for AI counseling. CA+ showed a 95.4% increase in average session length, from 14.38 to 28.10 minutes.

  • Problem

    AI counseling systems face difficulty sustaining long-term client engagement across evolving therapeutic contexts.

    Source: §1 · pp. 2–3

  • Method

    The paper proposes CA+, a cognition-augmented counselor agent framework for AI counseling.

    Source: §1 · p. 3, §8 · p. 43

  • Results

    CA+ showed a 95.4% increase in average session length, from 14.38 to 28.10 minutes.

    Source: §1 · p. 4

  • Takeaways & Limitations

    The authors report that CA+ improves client satisfaction and adherence to counseling standards.

    Source: §8 · p. 43

    The three-day study leaves the effectiveness of AI-driven counseling in sustained therapeutic contexts for longer-term research.

    Source: §7.4 · p. 43

Abstract

from arXiv · show

Current AI counseling systems struggle with maintaining effective long-term client engagement. Through formative research with counselors and a systematic literature review, we identified five key design considerations for AI counseling interactions. Based on these insights, we propose CA+, a Cognition Augmented counselor framework enhancing contextual understanding through three components: (1) Therapy Strategies Module: Implements hierarchical Goals-Session-Action planning with bidirectional adaptation based on client feedback; (2) Communication Form Module: Orchestrates parallel guidance and empathy pathways for balanced therapeutic progress and emotional resonance; (3) Information Management: Utilizes client profile and therapeutic knowledge databases for dynamic, context-aware interventions. A three-day longitudinal study with 24 clients demonstrates CA+'s significant improvements in client engagement, perceived empathy, and overall satisfaction compared to a baseline system. Besides, two licensed counselors confirm its high professionalism. Our research demonstrates the potential for enhancing LLM engagement in psychological counseling dialogues through cognitive theory, which may inspire further innovations in computational interaction in the future.

1 INTRODUCTION

CA+ targets the difficulty of sustaining long-term, context-aware counseling interactions by combining counselor-informed design considerations with a cognition-augmented framework. Its three modules coordinate therapeutic planning, communication, and information retrieval; the reported study found improved engagement and counseling relationship ratings compared with baseline.

1 INTRODUCTION

  • Interviews with counselors and a thematic review of AI counseling systems produced five design considerations that guided CA+ development.These include cognition-augmented hierarchical thinking, implicit client profiling, book-style data generation and retrieval, adaptive empathy, and adapting to AI-specific client expectations.Source: §1 · p. 3
  • CA+ is designed to maintain multi-turn contextual awareness and provide personalized, professionally aligned support over extended periods.The framework applies cognition-augmented reasoning to an LLM-based counselor agent.Source: §1 · p. 3
  • A three-day interaction study with 24 clients compared CA+ with a baseline system, while two licensed counselors assessed its professionalism and adherence to counseling standards.The counselors' assessment is described as validating CA+'s adherence to professional counseling standards.Source: §1 · pp. 3–4
  • CA+ combines hierarchical therapeutic planning, parallel guidance and empathy pathways, and client records with structured therapeutic knowledge for retrieval.The Therapy Strategies Module adapts through bidirectional planning; the Communication Form Module balances therapeutic effectiveness with emotional resonance.Source: §1 · p. 3
  • The paper offers design implications and guidelines for integrating LLMs into mental health support systems, based on system development and empirical findings.Its contribution list also identifies a controlled study showing improvements in engagement and satisfaction.Source: §1 · p. 4
  • 95.4% increase in average session length (14.38 to 28.10 minutes, p<0.001) and 46.7% improvement in overall counseling relationship ratings versus baseline.The paper also reports significant improvements across counseling quality metrics, client engagement indicators, and user feedback.Source: §1 · pp. 3–4

2 RELATED WORK

Prior work has advanced AI-assisted counseling, but the literature describes ongoing limits in long-term context, personalized adaptation, professional competence, and emotional responsiveness. Existing approaches address parts of these challenges, while the reviewed work identifies a need to integrate cognitive, emotional, and adaptive capabilities.

2.1 Evolution and Promise of AI-Assisted Mental Health Interventions

  • Rule-based systems offered structured therapeutic flows but could lack flexibility, nuance, and empathy; newer AI systems still face reliability, emotional intelligence, and long-term context limitations.The review notes that AI counseling remains largely in research and development rather than ready for widespread practical implementation in mental health settings.Source: §2.1 · p. 4

2.2 User Engagement in AI Psychological Therapy

  • Engagement challenges include shallow understanding of users over time, statelessness that impedes tracking changing user states, generic feedback, and difficulty establishing trust and emotional connection.The review links continuity across interactions to effective therapeutic relationships and describes individualized motivational feedback as an alternative to standardized encouragement.Source: §2.2 · p. 5

2.3 Approaches in AI-Assisted Counseling for Enhanced Engagement and Effectiveness

  • Cognitive architectures, expert knowledge, personalization, and empathetic agents address distinct needs, yet systems still struggle to combine logical decisions, sympathetic responses, and sustained relationships.The review calls for integrating cognitive processing, emotional modeling, and adaptive interaction techniques to better reflect the multifaceted expertise of human therapists.Source: §2.3 · p. 6

3 FORMATIVE RESEARCH: UNDERSTANDING HUMAN AND AI COUNSELING ENGAGEMENT

The formative research combines counselor practice and a review of AI counseling systems to understand the demands of effective counseling engagement. It identifies gaps and derives design considerations concerning adaptive therapy planning, client understanding, therapeutic knowledge, and empathic, trustworthy communication.

3.1 Formative Investigation of Human Counselor Practices

  • The investigation examined therapeutic procedures, strategic self-disclosure and client engagement, and professional documentation to inform AI design requirements.The formative study aimed to understand counselors’ expertise, cognitive processes, decision-making, and client engagement.Source: §3.1 · pp. 6–7

3.1.1 Methods and Participants.

  • Five licensed counselors, including one supervising counselor and four practicing counselors, participated in semi-structured interviews about their therapeutic approaches and professional practices.Interviews lasted approximately 60 minutes; transcripts were coded by one researcher, independently reviewed by a second, and disagreements resolved with a third.Source: §3.1.1 · p. 7
  • The analysis primarily focused on CBT and psychodynamic approaches in the Chinese therapeutic context, while the authors note that identified patterns may extend to other modalities.The analysis identified components of counselors’ standard operating procedures for the following section.Source: §3.1.1 · p. 7

3.1.2 Understanding Therapeutic Procedures.

  • Counselors organize therapy in three interconnected layers, from overarching phases to session agendas and moment-to-moment actions, forming a structured yet adaptable recursive planning system.Client responses at the action level can prompt session or goal changes, maintaining continuity while adapting to evolving needs.Source: §3.1.2 · pp. 7, 9
  • Within sessions, counselors select and deliver therapeutic actions based on client information, then reflect on responses and adapt when an action is ineffective.The process integrates current verbal content, prior discussion, non-verbal cues, and therapeutic goals.Source: §3.1.2 · pp. 8–9

3.1.3 Therapeutic Self-disclosure and Personalized Empathic Engagement.

  • Counselors strategically use self-disclosure to establish an authentic and empathetic connection with clients.The subsection also addresses personalized empathic engagement.Source: §3.1.2 · p. 9

3.1.2 Understanding Therapeutic Procedures

  • Counselors tailor empathy to client preferences and backgrounds, often creating emotional connection before moving to action steps.Self-disclosure can normalize client experiences while maintaining appropriate boundaries.Source: §3.1.2 · p. 9
  • Counselors record client statements, emotional responses, interventions, and significant developments, plan upcoming sessions around prior interactions and potential challenges, and use case conceptualizations to identify strengths, challenges, and treatment directions.Source: §3.1.2 · pp. 9–10
  • These documentation practices form a cognitive scaffold: session records inform case conceptualizations, which shape agendas in a feedback loop supporting responsive, personalized progression.The source describes the loop as enabling therapeutic continuity while accommodating emerging client needs.Source: §3.1.2 · p. 10

3.2 Formative Review of AI Counseling Approaches

  • The review analyzes current AI counseling interaction patterns and implementations alongside human counseling insights to ground the design of LLM-based counseling frameworks.Source: §3.2 · p. 10

3.2.1 Review Methodology

  • The authors followed PRISMA, searched databases from April to May 2024, and selected 55 papers for inclusion in the review.The initial queries returned 2,963 entries, and the review used screening and quality appraisal before selecting the final corpus.Source: §3.2.1 · pp. 10, 12

3.2 Formative Review of AI Counseling Approaches

  • The review identifies gaps in case conceptualization, treatment planning, therapeutic techniques, empathy, personalization, autonomy, alliance-building, and transparency in AI counseling.It also notes limitations in agent character design and psychoeducational delivery.Source: §3.2.2 · p. 12
  • Decision-tree and scripted systems follow rigid pathways that limit adaptive cognitive processing, while LLM counseling applications still show planning deficiencies despite mimicking counselor communication.The review describes the adaptive processing constraint as essential to effective counseling.Source: §3.2.2 · p. 12

3.2.1 Review Methodology

  • Current AI counseling systems often process turns in isolation and lack coherent therapeutic progression across interactions.This differs from human counselors’ coordination of session-level goals and moment-to-moment interventions based on evolving client needs.Source: §3.2.2 · p. 12
  • Prior AI counseling systems have made progress in empathy and therapeutic alliances, but typically address relationship components separately rather than offer a comprehensive therapeutic persona.The research argues for a comprehensive approach to relationship-building alongside cognitive and knowledge-processing capabilities.Source: §3.2.2 · p. 13

3.3 Design Considerations from Formative Research

  • The paper focuses on emulating therapeutic interactions, leaving safety protocols, data protection, ethical guidelines, and comprehensive risk mitigation for future research.Source: §3.3 · p. 13
  • DC4 emphasizes an empathetic therapeutic persona and relationship-building, while DC5 uses AI’s computational advantages for knowledge support and psychoeducation.Together, these considerations balance authentic therapeutic relationships with technological advantages.Source: §3.3 · p. 13

3.3.1 DC1: Cognitive Formulation of Counseling Processes

  • Counseling requires hierarchical goal, session, and action planning that adapts bidirectionally to higher-level aims and client feedback.The framework maintains overarching goals while flexibly responding to emerging client needs.Source: §3.3 · pp. 13–14
  • DC1 also calls for systematic reflection on current and historical interactions so AI counseling can refine therapeutic reasoning through experience.Reflection supports identifying biases or ineffective approaches and continuously refining strategies based on accumulated experience.Source: §3.3 · p. 14

3.3.2 DC2: Proactive and Continuous Client Psychological Profiling

  • DC2 requires proactive exploration and validation of client circumstances alongside continuous refinement of psychological profiles as new insights emerge.The dynamic profiles synthesize immediate concerns and evolving therapeutic needs to support more precise, aligned interventions.Source: §3.3 · pp. 14–15

3.3.3 DC3: Actionable Knowledge Base for Agent Intervention

  • DC3 calls for preprocessing and structuring therapeutic knowledge into standardized, retrievable operational formats that support real-time interventions.The design consideration addresses the gap between knowledge being accessible and its effective translation into practical interventions.Source: §3.3 · p. 15

3.3 Design Considerations from Formative Research

  • Therapeutic knowledge should be retrieved when conversation cues indicate a need, with selected components aligned to current goals and dynamics.The passage stresses that timing an intervention matters as much as the intervention itself.Source: §3.3 · p. 15
  • Retrieved knowledge should be personalized using client profiles and session history, with intervention intensity adjusted to readiness while preserving the therapeutic alliance.Readiness can be assessed through engagement and emotional stability; techniques may need modification for different clients.Source: §3.3 · p. 15

DC4: Therapeutic Persona Design for Authentic Empathy

  • AI counselors should establish a consistent, credible persona that conveys professionalism and empathy while transparently communicating their AI identity and limitations.Consistent behavior and clear capability statements can help build trust.Source: §3.3 · p. 16
  • AI-appropriate empathetic responses should acknowledge client emotions and convey understanding without merely mimicking human behavior.The responses should respect the boundaries of AI capabilities.Source: §3.3 · p. 16
  • Sustained alliance building requires emotional congruence, positive regard, and adapting communication tone and timing to client feedback.The goal is a sustainable relationship that combines human-like empathy with transparent recognition of AI capabilities and limitations.Source: §3.3 · p. 16

DC5: Meeting Stereotypical Expectations Toward AI Agents

  • Effective AI counseling should accommodate and leverage AI-specific expectations rather than only imitate human counselors; acknowledging non-human status may enhance effectiveness.The passage identifies clients’ expectations for analytical information and a subordinate service role as examples of these differences.Source: §3.3 · p. 16
  • Clients may expect AI counselors to provide precise, evidence-based information and to occupy a subordinate service role, shaping how they communicate.The cited literature describes expectations for more detailed information than human counselors provide and more directive user communication with AI.Source: §3.3 · p. 16
  • Compared with human counselors, users report significantly less concern about “bothering” AI, and may bypass social-distance barriers faster, sharing sensitive information within the first few interactions.Source: §3.3 · p. 16

4 CA+ AGENT FRAMEWORK

CA+ frames counseling as a recursive decision process that adapts therapeutic strategies to client feedback and contextual cues. Its planning and communication modules are supported by information management for real-time counseling operations.

4 CA+ AGENT FRAMEWORK

  • The framework was refined with participant feedback and professional oversight to address technical challenges and interaction requirements in AI-facilitated counseling.Source: §4 · p. 17

4.1 Framework Architecture and Module Design

  • CA+ uses a recursive, self-evolving decision graph that draws on therapeutic knowledge and professional counseling patterns to navigate problem-solving scenarios.Source: §4.1 · p. 17
  • The framework combines hierarchical therapy planning, parallel guidance and empathy strategies, and client and therapeutic-knowledge databases for real-time operations.The databases support dynamic decision-making and retrieval of structured knowledge during counseling sessions.Source: §4.1 · p. 17

4.2 DC1: Cognition Augmented Hierarchical Thinking

  • For interconnected concerns, the recursive architecture can address immediate stress while updating its understanding of underlying issues and longer-term goals.The example traces a client’s work stress to work-life balance, career anxiety, and family expectations.Source: §4.2 · p. 18

4.2.1 Recursive Layer Architecture.

  • The architecture separates overarching goals, session plans, and concrete action sequences, updating each layer only when its specified conditions are met.Goals change after persistent patterns or significant shifts; session plans adjust when accumulated feedback signals deviation from expected progress.Source: §4.2.1 · p. 18

4.2.2 Bidirectional Planning Mechanism.

  • Top-down planning derives session agendas and actions from therapy goals, while negative client feedback can propagate adjustments upward from actions to sessions and goals.Goal-level changes occur when deeper changes are required; otherwise, each layer can recalibrate based on feedback.Source: §4.2.2 · p. 19

Appendix B.1 for more comprehensive information.

This appendix details CA+’s design considerations and refinement process, including expectation-tailored profiling, knowledge retrieval, empathy and personalization, and prompts, alongside professional oversight and iterative changes addressing factual discrepancies and interaction challenges. It also describes shorter sessions and more actionable responses produced through refinement.

4.3 DC2: Conceptualization-Driven Implicit Client Profiling

  • CA+ builds and updates client profiles through therapeutic dialogue, using profile completeness and documented progress to adapt therapeutic progression.Its implicit profiling is guided by professional documentation practices.Source: §4.3 · pp. 19, 21

4.4 DC3: Book-style Data Generation and Retrieval

  • CA+ distills counseling texts into structured guidance and retrieves interventions using conversation context, therapeutic approach, and client profile to tailor responses.The knowledge base organizes extracted material with source hierarchy, functional tags, instruction-example pairs, and retrieval metadata.Source: §4.4 · pp. 21–22

4.5 DC4: Adaptive Empathy and Ecological Self

  • DC4 combines real-time empathy that adapts to emotional signals with long-term personalization and an ecological self integrating professional competence and personalized traits.The system can shift from empathetic reflection toward deeper exploration as emotional state improves, while updating strategies from historical profiles and records.Source: §4.5 · pp. 22–23

4.6 DC5: Adapting to AI-Specific Client Expectations

  • Prompts address AI-specific expectations through detailed evidence-based guidance, a deferential service role, reduced social burden, and boundaries for early intimacy and disclosure.They account for clients’ tendency to share personal information more readily with AI systems.Source: §4.6 · p. 24
  • The framework leveraged clients’ AI-specific expectations rather than fighting against them.Prior research suggested that acknowledging AI’s non-human status may enhance rather than detract from therapeutic effectiveness.Source: §4.6 · p. 24

4.7 System Implementation and Refinement

  • In a supervised preliminary study, 18 participants completed three-day counseling sessions in six batches, with reviews and improvements between batches.Licensed counselors and research team members monitored interactions in real time.Source: §4.7 · p. 25
  • Factual discrepancies fell from approximately 2.7% of first-batch responses to under 0.2% by the final batch, with no harmful therapeutic advice detected across 27 complete sessions.The team improved verified-resource retrieval, added guardrails around statistical claims, and used explicit uncertainty when evidence was weak.Source: §4.7 · p. 25
  • Client impatience with lengthy sessions led the team to shorten and modularize the agenda and streamline follow-up questions and clarifications.The changes aimed to maintain engagement and create shorter feedback loops; responses were also made more comprehensive, clear, and actionable.Source: §4.7 · p. 25

5 STUDY 1: COMPARATIVE CLIENTS EXPERIENCE STUDY

This three-day comparative study evaluates full CA+ against an ablated baseline across automated performance and client experience, using quantitative, qualitative, and behavioral measures. It examines counseling quality, empathy, self-disclosure, usability, engagement, conversational continuity, personalization, and clients’ perceptions of adaptive responses.

5.1 Participants

  • Of 28 recruits, four were excluded for data-quality concerns, leaving 24 consenting participants aged 20–44 (M = 25.58, SD = 5.24).Source: §5.1 · p. 26

5.2 Procedures

  • In a randomized between-subjects design, participants used full CA+ or a five-consideration-ablated baseline for 10 minutes daily over three days.Both conditions used the same model and interface, isolating the effects of the design considerations.Source: §5.2 · p. 27

5.3 Measures

  • Evaluation combined automated dialogue analysis, post-session client questionnaires, behavioral-log metrics, and interviews analyzed thematically.Source: §5.3 · pp. 27–28

5.4.1 System Performance.

  • CA+ outperformed the baseline system in therapeutic skills, including counseling-plan design (M_CA+ = 5.19, M_Baseline = 4.00; t(72) = 5.01, p = 0.000005 < 0.001) and proactive assessment (t(72) = 7.75, p = 8.17 × 10−11 < 0.001).Source: §5.4.1 · p. 28
  • CA+ exceeded Baseline on empathy (M_CA+ = 5.83, M_Baseline = 3.86) and follow-up ability (M_CA+ = 3.14, M_Baseline = 2.56).Source: §5.4.1 · p. 29
  • CA+ outperformed Baseline on memory (M_CA+ = 6.00, M_Baseline = 3.56) and personalization (M_CA+ = 3.39, M_Baseline = 2.86), though personalization was marginally significant.Source: §5.4.1 · p. 29
  • CA+ received more positive-feedback ratings than Baseline (M_CA+ = 4.36, M_Baseline = 2.47).Source: §5.4.1 · p. 29
  • CA+ received significantly higher Session Rating Scale ratings than the baseline across all evaluated dimensions.The dimensions included counseling relationship, goals and topics discussed, approach or method, and overall assessment.Source: §5.4.1 · p. 29, Figure 9 · p. 30, §5.4.2 · p. 30

5.4.2 Perceived Counseling Quality.

  • CA+ significantly outperformed the baseline across all four perceived-quality measures, including counseling quality, empathy, self-disclosure, and usability.The evaluation used the SRS, PETS, Self-Disclosure Scale, and SUS.Source: §5.4.2 · p. 30
  • Clients rated CA+ higher across PETS empathy dimensions and reported greater willingness to disclose personal experiences than with the baseline.The reported dimensions included emotional intelligence, empathy, and understanding clients’ goals; disclosure included comfort discussing experiences, ease talking about themselves, and honest expression.Source: §5.4.2 · p. 30
  • CA+ achieved a SUS mean of 80.88 versus 68.31 for the baseline, a 12.57-point difference described as “Excellent” usability.Source: §5.4.2 · p. 31

5.4.3 Client Behavioral Engagement.

  • CA+ clients provided longer replies than baseline (MCA+ = 30.70, SＤCA+ = 13.67; MBaseline = 22.24, SDBaseline = 11.29; t(71) = 2.86, p = 0.0062).CA+ session lengths increased from 26.98 minutes on day one to 29.75 minutes by day three; baseline lengths rose from 13.03 to 15.09 minutes.Source: §5.4.3 · pp. 32–33

5.4.4 Feedback on System Features.

  • CA+ was rated significantly higher than Baseline for guiding clients to take initiative (5.31 vs. 3.73) and keeping conversations centered on counseling (6.14 vs. 4.76).Source: §5.4.4 · p. 34
  • Client feedback highlighted CA+’s adaptive reasoning approach, which likely contributed to its significantly higher ratings in multi-turn understanding and reasoning.Source: §5.4.4 · p. 35
  • CA+ scored higher than the baseline for remembering client information (MCA+ = 5.83, SDCA+ = 1.62; MBaseline = 3.68, SDBaseline = 1.74).The difference was significant (t(72) = 5.48, p = 8.52 × 10−7 < 0.001).Source: §5.4.4 · p. 35
  • CA+ scored higher than the baseline for maintaining conversation continuity (MCA+ = 6.06, SDCA+ = 1.13; MBaseline = 3.51, SDBaseline = 1.70).The difference was significant (t(72) = 7.54, p = 2.21 × 10−10 < 0.001).Source: §5.4.4 · p. 35
  • CA+ scored higher than the baseline for learning and improvement (MCA+ = 5.56, SDCA+ = 1.23; MBaseline = 4.46, SDBaseline = 1.57) and personalized intervention plans (MCA+ = 5.47, SDCA+ = 1.12; MBaseline = 4.38, SDBaseline = 1.67).The reported differences were significant for both ratings (t(72) = 3.32, p < 0.01; t(72) = 3.30, p < 0.01, respectively).Source: §5.4.4 · p. 35
  • Clients reported significantly higher positive emotions, feeling valued and accepted, and self-disclosure with CA+ than with the baseline.Clients also described CA+’s personalized, empathetic approach as fostering a more engaging and positive experience.Source: §5.4.4 · p. 35
  • Clients valued CA+’s practical, tailored advice and its structured combination of theory and tasks for addressing problems.Reported examples included effective advice clients tried in real life and guidance that paired theories for perceiving and accepting emotions with specific tasks.Source: §5.4.4 · p. 35

6 STUDY 2: CERTIFIED COUNSELORS QUALITY ASSESSMENT

Two licensed counselors assessed CA+ using Study 1 transcripts, providing a professional evaluation complementary to client findings. They rated CA+ above baseline and described stronger session continuity and actionable consolidation.

6.1 Methodology

  • Counselors reviewed six balanced Study 1 transcripts, used the PAAC scale and semi-structured interviews, and analyzed transcripts thematically.Source: §6 · p. 36, §6.1 · p. 36

6.2 Findings

  • The expert evaluation found differences between CA+ and baseline across five dimensions, including system performance, session management, therapeutic approach, client outcomes, and potential improvements.Source: §6.2 · p. 36

6.2.1 Overall Assessment.

  • Counselors described CA+ as engaging in reciprocal dialogue, while baseline interactions were characterized as fragmented, one-way output.Source: §6.2.1 · p. 36
  • CA+ scored 5.35/7 versus baseline’s 1.85/7 on system-specific features, and 4.22/5 versus 1.94/5 on the Counseling Competence Scale.Source: §6.2.1 · p. 37

6.2.2 Sessions Management.

  • Counselors noted that CA+ recalled prior conversations and consolidated sessions into actionable homework for clients.Source: §6.2.2 · p. 37

2. Communication Practice: [...]

CA+ used supportive counseling techniques and guided a client toward positive action and communication. Counselors also identified opportunities to tailor interventions and use more tactful language on sensitive topics.

2. Communication Practice: [...].

  • A counselor praised end-of-session summaries and homework emphasis as reinforcing what clients can do after counseling ends.Source: §2 · p. 37

6.2.3 Therapeutic Approach and Techniques.

  • CA+ demonstrated active listening, empathy, clarification, probing for positive perspectives, normalizing thoughts, positive-resource focus, and respect for client boundaries.Source: §6.2.3 · p. 37

6.2.4 Positive Outcome.

  • CA+ encouraged clients to take positive action and communicate with partners; one client reported expressing their opinions and needs to a partner.Source: §6.2.4 · pp. 38–39

6.2.5 Opportunities for Enhancement.

  • Counselors suggested a wider range of interventions tailored to client types and whether concerns involve real-life situations or psychological issues.Source: §6.2.5 · p. 39
  • Counselors recommended more delicate, indirect phrasing for sensitive topics to support navigation of complex emotional situations and the therapeutic alliance.Source: §6.2.5 · p. 39

7 DISCUSSION

The discussion examines CA+’s evaluations and performance relative to the baseline, connecting its strengths in emotional support, guidance, memory, and continuity to design implications for AI counseling. It also considers adaptive empathy and contextual continuity, and identifies longer-term evaluation and improved handling of implicit meanings and silences as areas for further work.

7.1 Engaging Clients with CA+

  • For guiding client initiative, CA+ scored 5.31 (SD=1.33) versus Baseline’s 3.73 (SD=2.00), a significant difference (t(72)=3.98, p=0.000212).These ratings illustrate the broader finding that CA+ scored significantly higher than Baseline in client engagement and therapeutic quality.Source: §7.1 · p. 39
  • Clients rated CA+ higher in counseling quality, empathy, and willingness for self-disclosure, while counselors praised its session management and continuity across sessions.Counselors highlighted probing and clarification questions, empathetic responses, and goal-setting strategies.Source: §7.1 · pp. 39–40

7.2 Leveraging the Unique Features of AI in Therapeutic Interactions

  • After three days of consultation, participants perceived CA+ as having significantly higher Emotional Intelligence than Baseline, and experts praised its emotional acceptance and empathy.Counselors observed that its appreciation, encouragement, and intimacy could, to some extent, exceed typical human-to-human interaction boundaries.Source: §7.2 · p. 40
  • Users preferred detailed therapeutic guidance, and CA+ significantly outperformed Baseline in facilitating proactive communication (p<0.001).The discussion attributes this to integrating Book-style Data Generation and Retrieval (DC3) with Conceptualization-Driven Implicit Client Profiling (DC2) to produce information-dense responses.Source: §7.2 · p. 40
  • CA+ received significantly higher ratings than Baseline for client information recall and conversation continuity (p<0.001 for both measures).Participants noted that it remembered content from previous days.Source: §7.2 · p. 40
  • CA+ achieved higher personalized-learning scores (p<0.01) by balancing empathetic response patterns with goal tracking and adaptive therapy planning.The discussion frames empathy and proactive, goal-oriented guidance as complementary user expectations for AI counselors.Source: §7.2 · p. 40

7.3 Design Implications

  • Design implications favor iterative, multi-turn dialogue over immediate-response counseling, with progressive self-disclosure and context tracking supporting deeper exploration and continuity.Although long-term insight depth was not quantified, qualitative feedback indicated clients could better articulate inner experiences; preliminary evaluations also showed increased continuity and self-disclosure depth.Source: §7.3 · p. 41
  • The authors propose voice input and vocal-cue sentiment analysis to enrich contextual understanding, noting these enhancements remain experimental and need more comprehensive evaluation.They also report that follow-up questions and consistent context tracking are associated with increased client trust and satisfaction.Source: §7.3 · pp. 41–42
  • All reported improvements in multi-turn cognition, memory retention, and conversation continuity over baseline reached p<0.001.The discussion presents these quantitative results as evidence for CA+’s session-management design.Source: §7.3.2 · p. 42
  • By tracking context across turns and sessions, CA+ can naturally revisit topics, avoid repetitive questions, and build on established rapport.Participants and professional evaluators also valued its evolving responses, coherent cross-session dialogue, summaries, and follow-up tasks.Source: §7.3.2 · p. 42
  • Future iterations could broaden intervention techniques for diverse needs and distinguish situational challenges from underlying psychological issues.The discussion also calls for refining positive guidance to acknowledge the distinctive nature of human–AI therapeutic relationships.Source: §7.3.2 · p. 42, §7.3.3 · p. 43
  • CA+’s adaptive empathy was associated with clients feeling valued and accepted, increased self-disclosure versus baseline, and improved post-session emotional well-being.The discussion links emotional connection enabled by its ecological self-representation to deeper self-disclosure.Source: §7.3.3 · p. 42
  • Balancing adaptive empathy with AI-specific expectations supports judgment-free, consistently available interactions while maintaining therapeutic emotional connection.Professional evaluators endorsed CA+’s acceptance of clients’ emotions and normalization of their thinking patterns.Source: §7.3.3 · pp. 42–43

7.4 Limitations and Future Work

  • Because the study lasted three days, longer evaluations over weeks or months are needed to assess sustained engagement and therapeutic relationships.Future work should improve dialogue management, adaptive learning based on client progress, and tracking of changing needs and moods.Source: §7.4 · p. 43
  • Current AI struggles to interpret implicit meanings and silences, including topic avoidance, emotional reactions, and temporal cues in counseling dialogue.Future systems could use counseling-specific fine-tuning, specialized multi-agent interpretation, and multimodal interaction capabilities.Source: §7.4 · p. 43

8 CONCLUSION

CA+ is presented as a framework for addressing engagement challenges in AI counseling through contextual understanding, professional standards, and personalized interaction. The paper reports improved client satisfaction and counseling-standard adherence, and frames the framework as a foundation for future AI-augmented mental health research.

8 CONCLUSION

  • CA+ integrates mechanisms for contextual understanding, professional standards, and personalized interactions to address engagement challenges in AI counseling.The conclusion describes the framework as a Cognition Augmented Counselor Agent intended to overcome critical engagement challenges.Source: §8 · p. 43
  • The authors report empirical improvements in client satisfaction and adherence to counseling standards.The conclusion characterizes these studies as evidence of CA+'s effectiveness.Source: §8 · p. 43
  • CA+ is presented as a scalable foundation for future AI-augmented mental health research and broader access to psychological support.The conclusion links this potential to alleviating the global shortage of mental health professionals.Source: §8 · p. 44

A.1 Clinical Practice of Psychological Counseling

The clinical-practice interview guide asks counselors about their workflows and therapeutic approaches, then covers client intake, assessment, treatment planning, and progress tracking. It also addresses administration, professional development, challenging cases, and ethical practice.

A.1 Clinical Practice of Psychological Counseling

  • Questions cover counselors’ general clinical workflows, therapeutic orientations, techniques, adapting approaches, and responding to plateaus or resistance.The guide also asks about professional development and evaluating therapeutic effectiveness.Source: §A.1 · p. 47
  • The guide asks how counselors conduct initial consultations, assess client needs, formulate treatment plans, and determine suitability or referral.It also asks about setting therapeutic goals with clients.Source: §A.1 · p. 47
  • Questions address structuring ongoing sessions and tracking client progress throughout therapy.These topics are framed alongside the broader therapy process.Source: §A.1 · p. 47
  • Counselors are asked about supervision, continuous learning, and evaluating their own effectiveness.Source: §A.1 · pp. 47–48
  • The interview covers administrative procedures, documentation standards, scheduling and termination, and ethical guidelines and boundaries.Source: §A.1 · p. 48
  • The guide asks about crisis intervention, challenging cases, professional boundaries, and counselor self-care.Source: §A.1 · p. 48

A.2 Clinical Supervision in Psychotherapy

The clinical-supervision interview guide covers individual supervision sessions and the longer-term supervisory relationship. Its questions span session workflows, case materials, matching and arrangements, feedback, and competency assessment.

A.2 Clinical Supervision in Psychotherapy

  • The guide asks how supervisors find supervisees, select therapists, assess compatibility, and determine supervision frequency and duration.Source: §A.2 · p. 48
  • It also asks how supervisors evaluate supervision effectiveness, support counselors at different levels, and assess supervisor competencies.Source: §A.2 · p. 48
  • Questions about individual supervision address session workflows, handling advance case materials, and adapting supervision to therapist experience or relationship stage.Source: §A.2 · p. 48

B.1 DC1: Cognition Augmented Hierarchical Thinking

The appendix describes DC1's hierarchical prompt architecture for recursive therapeutic planning and DC2's documentation management prompts. Together, these descriptions cover multi-level navigation and documentation-linked case conceptualization and therapy phases.

B.1 DC1: Cognition Augmented Hierarchical Thinking

  • DC1 uses a multi-layered prompt architecture to recursively navigate therapeutic planning across levels of abstraction.The work-stress example addresses immediate concerns while updating understanding of underlying issues across temporal and conceptual levels.Source: §B.1 · p. 48, Figure 14 · p. 49

B.2 DC2: Conceptualization-Driven Implicit Client Profiling

  • DC2 prompts the system to update a client’s case conceptualization from conversation history and produce professional session records.The figure is identified as illustrating DC2's documentation-management prompt structures.Source: Figure 15 · p. 50, §B.2 · p. 50
  • DC2 links therapeutic phases to documentation completeness: the initial phase fills profile gaps, the middle phase enables targeted interventions, and the ending phase follows substantial documented progress.Source: §B.2 · p. 50

B.3 DC3: Book-style Data Generation and Retrieval

The appendix describes CA+’s counseling-knowledge processing, adaptive personalization and therapeutic persona, user interface, and system-feature evaluation scale.

B.3 DC3: Book-style Data Generation and Retrieval

  • DC3 converts professional counseling literature into tagged instruction-example pairs, then retrieves context-matched guidance using client history, case conceptualization, and session context.The generated pairs carry metadata for therapeutic approach, stage, and function; retrieval returns structured instructions and conversational examples.Source: §B.3 · pp. 50, 52

B.4 DC4: Adaptive Empathy and Ecological Self

  • DC4 maintains client preference profiles and monitors real-time receptiveness, adjusting interventions when resistance or emotional misalignment is detected.Profiles cover learning style, counseling preferences, topic sensitivities, and treatment boundaries.Source: §B.4 · p. 52
  • The Ecological Self prompt defines a consistent therapeutic persona through its values, communication characteristics, therapeutic qualities, and intervention style, with ethical boundaries.Source: §B.4 · p. 53

C USER INTERFACE

  • The interface provides account access, counseling-specific agent-role selection, and real-time chat designed for clarity and supportive exploration.Source: C USER INTERFACE · p. 54

D SELF-DESIGNED SCALE

  • The self-designed scale assesses continuity across sessions alongside memory, adaptive planning, listening, empathy, self-disclosure, engagement, therapy focus, reinforcement, and support.Source: D SELF-DESIGNED SCALE · p. 54

Related papers

Loading 2503.21365v1…